context(slice): out-of-range model selector is terminal instead of a recoverable per-call fault

nsaspy commented

2026-09-07 00:02:24 +00:00

Owner

Summary

A model-issued context(slice) request with a start index beyond the current payload length fails the entire run with context_operation_failed, instead of being surfaced as a recoverable per-call fault (the way malformed native tool-call arguments already are). The model cannot correct course after one bad slice guess.

Reproduction (live run, 2026-09-06)

Auto-Dig Prolog actor, native direct mode, pinned ref 39b278589dae8518e583778cd6671a6cb1e026c7 (with the Auto-Dig hotfix set applied). The model (z-ai/glm-5.3-flash via OpenRouter) had already made 6 model calls / 8 context calls when it requested:

{"operation": {"$term": "slice", "args": [6000, 2288]}}

Outcome error (rlm-result.json):

{
  "kind": "context_operation_failed",
  "message": "native context operation failed",
  "cause": {
    "kind": "out_of_range",
    "index": 6000,
    "message": "context selector is outside the payload",
    "operation": {"$term": "slice", "args": [6000, 2288]}
  }
}

The run terminated with exit 1 (phase=context kind=context_operation_failed). No retry, no repair attempt, no clamp.

Why this matters for Auto-Dig

The pinned runtime already treats recoverable native preflight faults, transient provider failures, and read-only tool errors as per-call events; an out-of-range selector is the same class: the model guessed an offset it could not verify.
A single bad slice guess should cost one call (ideally returning a bounded error/corrected-range hint the model can use), not the whole bounded research slice.
Observed twice in three live runs with identical config: run svc-20260906T235450Z (full success, no bad slices), run svc-20260906T235839Z (this failure), run svc-20260906T235958Z (full success again) — so it is model-behavior variance, not a deterministic request property.

Requested behavior

Out-of-range context(slice) selectors (and equivalent peek/search selectors) should be recoverable per-call faults: return a structured error the loop can feed back to the model (e.g. "payload length is N; valid range 0..N"), allow one bounded repair retry, and only fail the run if the model repeats the same invalid selector.
Alternatively clamp to the payload and record the clamped result in the trace, as long as the trace shows the requested vs. served range.

Environment

prolog-rlm pinned ref: 39b278589dae8518e583778cd6671a6cb1e026c7 + scripts/apply_prolog_rlm_hotfix.py patch set from starintel-gpt-auto-dig.
Runner: agents/auto_dig_rlm_runner.pl (context options max_bytes(262144), time_limit(5.0)).

## Summary A model-issued `context(slice)` request with a start index beyond the current payload length fails the **entire run** with `context_operation_failed`, instead of being surfaced as a recoverable per-call fault (the way malformed native tool-call arguments already are). The model cannot correct course after one bad slice guess. ## Reproduction (live run, 2026-09-06) Auto-Dig Prolog actor, native direct mode, pinned ref `39b278589dae8518e583778cd6671a6cb1e026c7` (with the Auto-Dig hotfix set applied). The model (z-ai/glm-5.3-flash via OpenRouter) had already made 6 model calls / 8 context calls when it requested: ```json {"operation": {"$term": "slice", "args": [6000, 2288]}} ``` Outcome error (`rlm-result.json`): ```json { "kind": "context_operation_failed", "message": "native context operation failed", "cause": { "kind": "out_of_range", "index": 6000, "message": "context selector is outside the payload", "operation": {"$term": "slice", "args": [6000, 2288]} } } ``` The run terminated with exit 1 (`phase=context kind=context_operation_failed`). No retry, no repair attempt, no clamp. ## Why this matters for Auto-Dig - The pinned runtime already treats recoverable native preflight faults, transient provider failures, and read-only tool errors as per-call events; an out-of-range selector is the same class: the model guessed an offset it could not verify. - A single bad slice guess should cost one call (ideally returning a bounded error/corrected-range hint the model can use), not the whole bounded research slice. - Observed twice in three live runs with identical config: run `svc-20260906T235450Z` (full success, no bad slices), run `svc-20260906T235839Z` (this failure), run `svc-20260906T235958Z` (full success again) — so it is model-behavior variance, not a deterministic request property. ## Requested behavior - Out-of-range `context(slice)` selectors (and equivalent peek/search selectors) should be recoverable per-call faults: return a structured error the loop can feed back to the model (e.g. "payload length is N; valid range 0..N"), allow one bounded repair retry, and only fail the run if the model repeats the same invalid selector. - Alternatively clamp to the payload and record the clamped result in the trace, as long as the trace shows the requested vs. served range. ## Environment - prolog-rlm pinned ref: `39b278589dae8518e583778cd6671a6cb1e026c7` + `scripts/apply_prolog_rlm_hotfix.py` patch set from starintel-gpt-auto-dig. - Runner: `agents/auto_dig_rlm_runner.pl` (context options `max_bytes(262144), time_limit(5.0)`).

nsaspy referenced this issue from nsaspy/starintel-gpt-auto-dig

2026-09-07 00:02:39 +00:00

auto-dig service: crash after successful run-branch push left a green run recorded as failed (fixed on test branch, needs landing) #2148

nsaspy referenced this issue from nsaspy/starintel-gpt-auto-dig

2026-09-07 02:22:49 +00:00

feat(auto-dig): read-only StarIntel corpus MCP tools + standalone service runner landing #2149

nsaspy referenced this issue

2026-09-08 01:29:17 +00:00

[ARADR][MACHINE-SPIRIT] 8 topics × 8 alternating research/design passes #387

nsaspy referenced this issue

2026-09-08 01:29:18 +00:00

[MACHINE-SPIRIT 7/8] Meta-reasoning and bounded self-improvement: failure learning, strategy evolution, rule promotion, self-model and expert synthesis #394